Using machine learning to predict estimated glomerular filtration rate in type 2 diabetes patients: A 4...

Area of Science:

• Nephrology
• Endocrinology
• Data Science

Background:

• Global rise in type 2 diabetes mellitus (T2D) and diabetic kidney disease (DKD).
• Need for accurate prediction of estimated glomerular filtration rate (eGFR) decline in T2D patients.

Purpose of the Study:

• Compare the predictive accuracy of four machine learning (Mach-L) methods against multiple linear regression (MLR).
• Identify and rank key risk factors for DKD.
• Assess the potential of Mach-L for early DKD risk stratification.

Main Methods:

• Utilized data from 907 T2D patients followed for 4 years (2013-2019).
• Applied four Mach-L methods: classification and regression tree, random forest, artificial neural network, and eXtreme Gradient Boosting.
• Used multiple linear regression (MLR) as a benchmark and Shapley additive explanation for model interpretability.

Main Results:

• Random forest, classification and regression tree, and eXtreme Gradient Boosting outperformed MLR in predicting eGFR.
• Top risk factors for eGFR decline: body mass index (BMI), HDL-C, urine microalbumin creatinine ratio (MCR), LDL-C, diabetes duration, and age.
• BMI was the most influential factor, followed by HDL-C, MCR, LDL-C, diabetes duration, and age.

Conclusions:

• Mach-L methods demonstrate superior accuracy over MLR for predicting eGFR in T2D patients.
• Identified key modifiable and non-modifiable risk factors for DKD.
• Highlights Mach-L's potential for enhancing early DKD risk stratification and guiding interventions to preserve renal function.